偏最小二乘-判别分析模型影响因素初步研究

Elementary research on the influence factors of the partial least square - discriminant analysis models

中文关键词:  红外光谱  判别模型  得分  小麦.

英文关键词:Infrared Spectroscopy  Discriminant Model  Score  Wheat.

中文摘要:

      目的 研究偏最小二乘-判别分析模型的影响因素。方法 以小麦为例, 采用人工加速老化结合近、中红外光谱技术采集陈化、非陈化小麦粉的近、中红外光谱, 分别针对小麦粉衰减全反射中红外数据、衰减全反射中红外一阶导数数据、近红外一阶导数数据以及小麦颗粒近红外一阶导数数据建立偏最小二乘-判别分析模型, 分别获取模型参数及第1、2主成分得分分布情况。结果 小麦粉近红外一阶导数建模的第1、2主成分得分散点图具有明显的两类样本分布。外部验证数据表明, 小麦粉近红外一阶导数所建模型具有更好的稳健性。结论 样品的物理形态、数据预处理以及红外光谱波段皆会对偏最小二乘-判别分析模型结果产生影响; 同时, 红外光谱偏最小二乘-判别分析模型的评价需要从正确识别率、模型的校正测定系数、交互验证测定系数、校正均方根误差、交互验证均方根误差以及模型主要成分得分分布等多种情况综合考虑。

英文摘要:

      Objective To investigate the influence factors of partial least square-discriminant analysis models. Methods Taking wheat as the example, the attenuated total reflectance-mid-infrared and near-infrared spectroscopic data of the artificial accelerated aging wheat and the normal control group were collected respectively. The partial least square-discriminant analysis models were developed by attenuated total reflectance-mid-infrared data of wheat flour, 1st-derivative attenuated total reflectance-mid-infrared data of wheat flour, 1st-derivative near-infrared data and 1st-derivative near-infrared data of wheat, and the score scatter plots of the first and the second principal components were acquired respectively. Results The score scatter plot of 1st-derivative near-infrared data of wheat flour had specific boundary of two groups. The external prediction data showed that 1st-derivative near-infrared data of wheat flour was the most robust model among the 4 conditions. Conclusion The physical form, data preprocessing and the band of infrared spectroscopy would influence the partial least square-discriminant analysis model. Meanwhile, the evaluation to partial least square-discriminant analysis model based on infrared spectroscopic data should consider the correct recognition rate, the parameters of determination coefficients of calibration and cross validation, root mean square error of calibration and cross validation and the main principal component’s score distribution synthetically.